Definition
Plain language
A way of treating an AI agent's task like source code and compiling it into an optimized plan before running anything.
As stated in the literature
A just-in-time compilation framework for computer-use agents that statically verifies state-flow contracts on cached tool plans, ranks candidate programs by a cost model, and picks a Monte-Carlo-simulated execution strategy at runtime.
Also called: JIT-Planner, JIT planner
Why it matters: Pure step-by-step LLM control is slow and error-prone for repetitive UI work, and compiling and caching plans makes computer-use agents both faster and more reliable.
For example, when asked to fill in a familiar form, Agent JIT reuses a verified plan from cache instead of letting the model rediscover each click from scratch.
Heard on the show
“It's called "Agent JIT Compilation for Latency-Optimizing Web Agent Planning and Scheduling," out of Stanford, and it takes that observation seriously.”Episode 063 — Why Web Agents Are Slow: A Compiler-Style Fix for Computer-Use Latency